1 /*
2  * e_expf.c - single-precision exp function
3  *
4  * Copyright (c) 2009-2018, Arm Limited.
5  * SPDX-License-Identifier: MIT
6  */
7 
8 /*
9  * Algorithm was once taken from Cody & Waite, but has been munged
10  * out of all recognition by SGT.
11  */
12 
13 #include <math.h>
14 #include <errno.h>
15 #include "math_private.h"
16 
17 float
expf(float X)18 expf(float X)
19 {
20   int N; float XN, g, Rg, Result;
21   unsigned ix = fai(X), edgecaseflag = 0;
22 
23   /*
24    * Handle infinities, NaNs and big numbers.
25    */
26   if (__builtin_expect((ix << 1) - 0x67000000 > 0x85500000 - 0x67000000, 0)) {
27     if (!(0x7f800000 & ~ix)) {
28       if (ix == 0xff800000)
29         return 0.0f;
30       else
31         return FLOAT_INFNAN(X);/* do the right thing with both kinds of NaN and with +inf */
32     } else if ((ix << 1) < 0x67000000) {
33       return 1.0f;               /* magnitude so small the answer can't be distinguished from 1 */
34     } else if ((ix << 1) > 0x85a00000) {
35       __set_errno(ERANGE);
36       if (ix & 0x80000000) {
37         return FLOAT_UNDERFLOW;
38       } else {
39         return FLOAT_OVERFLOW;
40       }
41     } else {
42       edgecaseflag = 1;
43     }
44   }
45 
46   /*
47    * Split the input into an integer multiple of log(2)/4, and a
48    * fractional part.
49    */
50   XN = X * (4.0f*1.4426950408889634074f);
51 #ifdef __TARGET_FPU_SOFTVFP
52   XN = _frnd(XN);
53   N = (int)XN;
54 #else
55   N = (int)(XN + (ix & 0x80000000 ? -0.5f : 0.5f));
56   XN = N;
57 #endif
58   g = (X - XN * 0x1.62ep-3F) - XN * 0x1.0bfbe8p-17F;  /* prec-and-a-half representation of log(2)/4 */
59 
60   /*
61    * Now we compute exp(X) in, conceptually, three parts:
62    *  - a pure power of two which we get from N>>2
63    *  - exp(g) for g in [-log(2)/8,+log(2)/8], which we compute
64    *    using a Remez-generated polynomial approximation
65    *  - exp(k*log(2)/4) (aka 2^(k/4)) for k in [0..3], which we
66    *    get from a lookup table in precision-and-a-half and
67    *    multiply by g.
68    *
69    * We gain a bit of extra precision by the fact that actually
70    * our polynomial approximation gives us exp(g)-1, and we add
71    * the 1 back on by tweaking the prec-and-a-half multiplication
72    * step.
73    *
74    * Coefficients generated by the command
75 
76 ./auxiliary/remez.jl --variable=g --suffix=f -- '-log(BigFloat(2))/8' '+log(BigFloat(2))/8' 3 0 '(expm1(x))/x'
77 
78   */
79   Rg = g * (
80             9.999999412829185331953781321128516523408059996430919985217971370689774264850229e-01f+g*(4.999999608551332693833317084753864837160947932961832943901913087652889900683833e-01f+g*(1.667292360203016574303631953046104769969440903672618034272397630620346717392378e-01f+g*(4.168230895653321517750133783431970715648192153539929404872173693978116154823859e-02f)))
81             );
82 
83   /*
84    * Do the table lookup and combine it with Rg, to get our final
85    * answer apart from the exponent.
86    */
87   {
88     static const float twotokover4top[4] = { 0x1p+0F, 0x1.306p+0F, 0x1.6ap+0F, 0x1.ae8p+0F };
89     static const float twotokover4bot[4] = { 0x0p+0F, 0x1.fc1464p-13F, 0x1.3cccfep-13F, 0x1.3f32b6p-13F };
90     static const float twotokover4all[4] = { 0x1p+0F, 0x1.306fep+0F, 0x1.6a09e6p+0F, 0x1.ae89fap+0F };
91     int index = (N & 3);
92     Rg = twotokover4top[index] + (twotokover4bot[index] + twotokover4all[index]*Rg);
93     N >>= 2;
94   }
95 
96   /*
97    * Combine the output exponent and mantissa, and return.
98    */
99   if (__builtin_expect(edgecaseflag, 0)) {
100     Result = fhex(((N/2) << 23) + 0x3f800000);
101     Result *= Rg;
102     Result *= fhex(((N-N/2) << 23) + 0x3f800000);
103     /*
104      * Step not mentioned in C&W: set errno reliably.
105      */
106     if (fai(Result) == 0)
107       return MATHERR_EXPF_UFL(Result);
108     if (fai(Result) == 0x7f800000)
109       return MATHERR_EXPF_OFL(Result);
110     return FLOAT_CHECKDENORM(Result);
111   } else {
112     Result = fhex(N * 8388608.0f + (float)0x3f800000);
113     Result *= Rg;
114   }
115 
116   return Result;
117 }
118